Renmin's CoworkOS Harness Lifts Qwen3.6 From 33.19 to 42.95
TL;DR
- CoworkOS, from Renmin University of China and five collaborators, claims performance competitive with Codex and DeepSeek Harness across 12 benchmarks and six application categories.
- A three-layer architecture separates strategy from execution: the model decides actions, delegation and acceptance while a runtime maintains state and feedback.
- Using GLM-5.3 as backbone, five rounds of RSI-driven post-training lifted Qwen3.6-35B-A3B's eight-benchmark average from 33.19 to 42.95, a +9.76 gain.
A new preprint led by Renmin University of China describes CoworkOS, an agent harness its authors call "highly competitive with frontier agent applications such as Codex and DeepSeek Harness" across 12 benchmarks and six categories of real-world applications.
The system targets what the paper calls "long-horizon work," using a three-layer architecture that separates strategy from execution. In the abstract's own words, "the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback." Humans and agents inspect and revise the same files, including "evolving documents, presentations, CAD models, and Blender scenes."
The headline experiment is a self-improvement run. GLM-5.3 serves as backbone to conduct "RSI-driven automated post-training" on Qwen3.6-35B-A3B, and over five rounds "the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76)." No per-benchmark breakdown appears in the abstract itself.
The write-up on Hugging Face lists co-authors from Fudan, Peking, Shanghai Jiao Tong, NUS and East China Normal alongside Renmin, and lands amid a steady run of agent-harness research we've been tracking: 526 agent stories in the last 90 days on our side.
Originally reported by huggingface.co
Read the original article →Original headline: Renmin University's CoworkOS Agent Harness Tops SOTA Across 4 of 12 Benchmarks, Lifts Qwen3.6-35B +29.4% via Self-Improvement